An estimation theory-based approach for speech enhancement

This contribution presents an efficient technique for the speech enhancement of a signal using statistical estimators which are based on squared magnitude spectra's. In any speech enhancement systems, an estimate of power spectral density is required. As conventional methods for noise elimination fails due to the non-stationary properties of the speech signal, in this context, minimum mean square error (MMSE) and maximum a posterior (MAP) estimators are derived based on Gaussian statistical model. The acquisition function which is obtained in the MAP estimator is same as the acquisition function used in the ideal binary masking. As a binary masking depends on the signal-to-noise ratio (SNR), if the SNR value exceeds 0 dB then the value assumes to be 1 otherwise 0. The results accomplished using the proposed estimator embarked with better enhancement of the speech signal than the standard minimum mean square error spectral power estimator, with low residual noise and low speech distortion.